Weakly-supervised deep learning for breast tumor segmentation in ultrasound images

Abstract

Breast tumor segmentation is a preliminary step for breast tumor diagnosis. It is expensive to obtain pixel-level labels for breast tumor segmentation. We aim to perform breast tumor segmentation in breast ultrasound (BUS) images with image-level labels. To this end, we propose a class activation mapping with level set (CAM-LS) method. Given a training dataset consisting of BUS images with or without breast tumors which naturally contain image-level labels, a classification network is trained for breast tumor recognition. Class activation maps (CAMs) are generated for positive predictions, which highlight discriminative regions of breast tumors. Anatomical constraints are used to reduce the search space for breast tumors. For breast tumor recognition, the proposed method achieves sensitivity of 92.0%, precision of 92.8%, specificity of 95.2%, accuracy of 93.9% and F1-score of 0.924. For breast tumor segmentation, the proposed method obtains Dice similarity coefficient (DSC) of 73.5 ± 18.0% and interaction-over-union (IoU) of 60.8 ± 19.7%. The proposed CAM-LS method performs automated breast tumor segmentation with image-level labels only and is demonstrated in the experiments to have good generalization on different ultrasound platforms.